Selected Publications

V2X-ViT: Vehicle-to-Everything Cooperative Perception with Vision Transformer
Runsheng Xu*, Hao Xiang*, Zhengzhong Tu*, Xin Xia, Ming-Hsuan Yang Jiaqi Ma
ECCV 2022 [Paper] [Code]
The first unified transformer architecture for V2X perception, which can capture the heterogeneity of V2X systems with strong robustness against various V2X noises.

OPV2V: An Open Benchmark Dataset and Fusion Pipeline for Perception with Vehicle-to-Vehicle Communication
Runsheng Xu*, Hao Xiang*, Xin Xia, Xu Han Jinlong Li Jiaqi Ma
ICRA 2022 [Project] [Paper] [Code]
The first large-scale open V2V perception dataset and the first open V2V coding framework.

TridentNetV2: Lightweight Graphical Global Plan Representations for Dynamic Trajectory Generation
David Paz, Hao Xiang, Andrew Liang, Henrik I. Christensen
ICRA 2022 [Paper]
Generate dynamic trajectories for autonomous navigation via nominal graph-based global plan and a lightweight representation without the dependence on HD maps

Probabilistic Semantic Mapping for Urban Autonomous Driving Applications
David Paz*, Hengyuan Zhang*, Qinru Li*, Hao Xiang*, Henrik I. Christensen
IROS 2020 [Paper]
Build probabilistic semantic maps by leveraging segmentation network uncertainty and the LiDAR intensity prior knowledge.

Selected Projects

Particle Filter SLAM and Textured Map [Paper]
Implemented Particle Filter SLAM from scratch for stereo cameras and LiDAR data. Colored the floor of the map by using 3D geometry.

Visual-Inertial SLAM [Paper]
Implemented EKF from scratch with IMU data and camera data to jointly estimate the position of landmarks and robot.

Search-based and sampling-based algorithms for 3D-Euclidean space [Paper]
Compare performance of weighted A* and RRT/RRT* algorithms for various testing scenes. Implemented collision detection algorithms between AABB and line segment.

Heuristic-guided Reinforcement Learning Algorithms for lunar lander [Paper]
Using heuristics for all stage of training would introduce human bias. Instead of using heuristic for the final decision process, we utilize heuristic as a coach for the early stage training. Once the model has gained enough positive samples, we use a decaying factor to reduce the bias of heuristics. In this way, the algorithm can learn much faster without introducing human bias.